Amazon recommendations

[cellio]

I assume that Amazon's recommendations are computed largely from overlapping purchases (the "people who bought X also bought Y" approach), though they are also influenced by personal data like what I've put on my wish list and, apparently, ratings. These methods are good at capturing positive data, but they're weak for negative data. And the similarity groupings seem to be broad (not deep), which sometimes means wacky results.

There seems to be no way to tell Amazon's engine that yes, I am interested in (Jewish) biblical texts and Hebrew, but this does *not* mean I'm interested in Greek or Christian texts. There seems to be no way to tell it that I'm (currently) uninterested in books about C++, though Java and general programming are interesting. There seems to be no way to stop getting it to offer me baroque, classical, and modern music (or books on music) just because I'm interested in medieval and renaissance music. I have no idea if the copious "not interested" indications actually feed into the data pool, or if they merely act as a filter on what to show on the list.

As long as I'm going on about the recommendations scheme: there are different flavors of "not interested", and there seems to be no way to capture that. I might be uninterested because of the subject -- e.g. don't offer me books about football. I might be uninterested because I own something else that fills the same niche -- a different, similar text on Hebrew, or a compilation album that contains most of the good stuff from the CD offered. In the latter case, I don't want it to draw broad conclusions about the genre/topic (she's not interested in Hebrew any more, or she no longer likes Eric Bogle, etc). If I check "not interested" in those cases, am I sabotaging myself down the road?

I do, by the way, sometimes send questions/suggestions like these to Amazon via their feedback form, but there is absolutely no way for me to tell if that's effective (or welcome).

Comments

[msmemory.livejournal.com]

When she was living, Mother ordered from Amazon through me. Not using a computer herself, she'd phone me and say "have them send me the following three books" then she'd reimburse me. Her tastes being different from mine, especially more religious and romantic, my Amazon recommendations have been really skewed. Presumably it'll wear off, just like Lands' End eventually stopped sending me kids' catalogs several years after I bought a shower gift through them.

[cellio]

Somewhere on the recommendations page is a link to a page where you can do some configuration. Specifically, for each item you've bought, you can check a box to exclude it from consideration in recommendations. That wasn't always there and gift purchases caused me some wacky recommendations (especially when buying stuff for my sister's kids), but that doesn't seem to be a problem now.

[tangerinpenguin.livejournal.com]

The page that cellio refers to is:

https://www.amazon.com/gp/yourstore/iyr

[sui66iy.livejournal.com]

Back in the day (circa dot.com boom) there was a brief fad for collaborative filtering technology. The idea was that we could automatically compute good recommendations using stuff like latent semantic analysis. I remember asking someone, "if the latent semantic analysis stuff is so hot, how come all the recommendation engine stuff on Amazon sucks so much?" They told me that, in practice, the good algorithms for doing this are too expensive to run, so in reality people implement pretty simple genre-based recommendations with pretty heavy input from the marketing guys about what they want to sell you today.

I have no idea if that claim is true (nor can I remember who told me). I think it's just as likely that automatic recommendation tech isn't all that great. But it's an interesting alternative possibility. I should ask Dom... he probably has some idea.

[gregbo.livejournal.com]

They told me that, in practice, the good algorithms for doing this are too expensive to run, so in reality people implement pretty simple genre-based recommendations with pretty heavy input from the marketing guys about what they want to sell you today.

I don't know for sure, but my guess is that you're correct. In any event, I'd rather have cheap(er) books and free shipping in exchange for a few negative hits on my preferences.

[hobbitblue.livejournal.com]

I'd like Amazon to realise that Latin != Greek, I'd be happier to be offered Hebrew to be honest. I'm glad they have the "why was this recommended to me" button now, because sometimes the ideas are a bit head-scratching "We recommnended *generic modern poppy music by band you've never heard of* because you bought a copy of Weight Training for Women!" What?!

recommendations

[brokengoose.livejournal.com]

A few other factors in the whole thing:

The recommendation feature is not critical. Therefore, it makes sense to keep the UI as simple as possible. Adding buttons for things like "I like things like this but I'm not interested in THIS item" may make the UI too confusing for ordinary users.

As a mostly book and music store, it's in their best interest if your interests keep expanding. That is, if given the choice to display a CD that they -know- you'll like or to display a CD that you -might- like, they might be better off in the long term to make the riskier recommendation, particularly if it looks like you're "just looking around".

There's also the issue of sample size, particularly with things like medieval music. It could well be that most people who are into medieval music are also into other classical music. Therefore, your "not interested" clicks are simply lost in the crowd of classical music fans who do like both genres.

[metahacker.livejournal.com]

I have a slightly inside view here, because I was interviewing with some of the people that do this. Basically they're making the algorithms as smart as possible for as dumb a consumer as possible. They read your suggestions, but it's back to the very hard problem of making an interface that is usable by everyone but useful for geeks. Barring an "expert" mode (which doesn't work, because EVERYONE thinks they're an expert, especially the ones that aren't), that's a tough problem.

They seem to have gone with 'least harmful conclusion', so I wouldn't say checking "not interested" means you'll never see XXX again; it's all taken into account as correlations, not as sharp causality.

...they're still hiring, of course, if someone you know has any ideas on how to do this better.